Goldman Sachs: AI Capital Expenditure to Reach $1.73 Trillion in 2026-2027

Goldman Sachs research indicates that the AI capital expenditure of major U.S. cloud service providers is expected to reach approximately $1.73 trillion in 2026-2027. Based on a 15% annualized return on invested capital (ROIC), six companies—Alphabet, Microsoft, Amazon, Meta, Oracle, and SpaceX—will need to generate about $1.42 trillion in cumulative revenue from 2028 to 2030 to meet the return requirements of this round of AI computing power investment, which translates to approximately $11.6 billion in revenue per gigawatt of computing power per year. Goldman Sachs divides AI computing power construction into three phases: approximately $633 billion in capital expenditure from 2023 to 2025; about $1.73 trillion from 2026 to 2027; and an expected increase to about $4.14 trillion from 2028 to 2030. In terms of demand, AWS, Azure, and Google Cloud had a combined backlog of approximately $1.69 trillion as of the second quarter of 2026, representing a year-on-year increase of about 152%. Goldman Sachs estimates that the three cloud service providers will have about $1.22 trillion in capital expenditure from 2026 to 2027, corresponding to approximately $1 trillion in revenue from 2028 to 2030 to meet the 15% ROIC threshold, which is only about 59% of the current backlog. Goldman Sachs believes that the current pressure on AI capital expenditure returns mainly stems from large-scale upfront investments and does not indicate a structural profitability issue in the AI economic model. Their calculations show that with an ROIC target of 0% to 30%, the cumulative revenue required from 2028 to 2030 is approximately $908 billion to $1.89 trillion. Additionally, Goldman Sachs expects AI infrastructure capital expenditure to be about $1.3 trillion in 2027, rising to about $2 trillion in 2028; the corresponding new deployment scale of AI data centers is expected to reach 35GW and 57GW, respectively. Goldman Sachs believes that the acceleration of enterprise AI applications from the experimental stage to actual deployment, along with the continuous growth of cloud service backlogs, will become important demand support for the monetization of AI computing power in the future.